Integrated sensing and communication (ISAC) and reconfigurable intelligent surfaces (RISs) are viewed as promising technologies for the future 6G wireless networks. ISAC designs assisted by RISs are particularly attractive for tracking and localization problems in internet of everything (IoE) applications. In particular, RISs can be deployed to realize a smart radio environment (SRE) that tracks the user equipment (UE) in blind spaces, i.e., where the direct line-of-sight (LoS) wireless link is not available. This paper proposes a deep learning framework to integrate RIS-assisted mobile UE localization and communication in the 60 GHz band. The number of RIS and their electronic steering angles are investigated to support both localization and communication processes implemented on shared time resources. The UE localization is obtained through DL algorithms based on convolutional neural networks (CNN) and vision transformers (ViT) structures. The proposed algorithms are trained using a wide variety of physical parameters such as number of RIS steering angles, RIS area size, and number of antenna at the base station (BS). The system performance is measured in terms of achieved positioning root mean squared error (RMSE), algorithm complexity, and inference time. A Cramér-Rao bound for estimating the localization error based on RISs deployment, is also provided. Localization accuracy, frame efficiency and throughput tradeoffs are explored for different IoT setups.

This paper considers an indoor smart radio environment (SRE) where a Base Station (BS) communicates with a set of user equipment (UE) in the sub-THz band by reconfigurable intelligent surfaces (RISs), as the direct BS-UE links are obstructed. Motivated by the sparsity of the sub-THz channel, we model each RIS as an electronically steerable reflector, which can be described by a single parameter, i.e., the steering angle. We focus on the case where the positions of the UEs are unknown and have to be estimated. Specifically, we propose a novel approach to use RISs for jointly localizing the UEs and optimizing the communication performance. The UE localization is made possible by a dedicated RIS and is handled by a machine learning (ML) algorithm, which exploits the signal transmitted by the UEs and received at the BS. Once the estimates of the UE positions are available, the downlink communication between BS and UEs is optimized by properly selecting the electronic steering angle of the RISs so as to maximize the network throughput. By numerical simulations, the paper shows how the system performance is affected by the area of the RISs, by the number of antennas available at the BS, and by the number of steering angles scanned by the RIS used for localization.